Direct Sparse Visual-Inertial Odometry using Dynamic Marginalization (1804.05625v1)

Abstract: We present VI-DSO, a novel approach for visual-inertial odometry, which jointly estimates camera poses and sparse scene geometry by minimizing photometric and IMU measurement errors in a combined energy functional. The visual part of the system performs a bundle-adjustment like optimization on a sparse set of points, but unlike key-point based systems it directly minimizes a photometric error. This makes it possible for the system to track not only corners, but any pixels with large enough intensity gradients. IMU information is accumulated between several frames using measurement preintegration, and is inserted into the optimization as an additional constraint between keyframes. We explicitly include scale and gravity direction into our model and jointly optimize them together with other variables such as poses. As the scale is often not immediately observable using IMU data this allows us to initialize our visual-inertial system with an arbitrary scale instead of having to delay the initialization until everything is observable. We perform partial marginalization of old variables so that updates can be computed in a reasonable time. In order to keep the system consistent we propose a novel strategy which we call "dynamic marginalization". This technique allows us to use partial marginalization even in cases where the initial scale estimate is far from the optimum. We evaluate our method on the challenging EuRoC dataset, showing that VI-DSO outperforms the state of the art.

• The paper introduces VI-DSO, a method that jointly minimizes photometric and IMU errors through dynamic marginalization.
• It integrates inertial preintegration with direct sparse imaging to improve initialization and accurately track camera motion.
• Evaluation on the EuRoC dataset confirms enhanced accuracy and robustness over traditional keypoint-based odometry systems.

Direct Sparse Visual-Inertial Odometry using Dynamic Marginalization

The paper by Stumberg, Usenko, and Cremers introduces VI-DSO, an innovative approach to visual-inertial odometry (VIO) that advances the accurate and efficient determination of camera poses and sparse scene geometry by jointly minimizing photometric and inertial measurement errors within a unified energy framework. This approach builds on Direct Sparse Odometry (DSO), extending it by integrating an Inertial Measurement Unit (IMU) to significantly improve motion estimation performance, especially in challenging environments. VI-DSO presents several contributions that enhance robustness and precision over existing state-of-the-art methods.

The core methodology centers on a combined energy functional that incorporates photometric errors and IMU constraints, allowing for synchronized refinement of 3D geometry and motion parameters. Unlike traditional keypoint-based odometry systems, this approach minimizes a photometric error, facilitating the tracking of diverse intensity-gradient pixels rather than limited features like corners. The IMU preintegration strategy aggregates measurements across multiple frames, furnishing additional constraints between keyframes, which are seamlessly incorporated into the optimization process. A salient feature is the explicit modeling of scale and gravity direction, enabling immediate initialization with arbitrary scale values rather than deferring until observability conditions are met. The use of dynamic marginalization tools ensures the system's consistency even when initial scale estimates deviate notably from optimal values.

Evaluation on the EuRoC dataset indicates that VI-DSO consistently surpasses existing algorithms in terms of accuracy, notably in environments characterized by poor illumination and significant motion blur. The notion of dynamic marginalization provides a mechanism for adaptive information reduction from older states while maintaining consistency across the system. This adaptability is crucial for maintaining operational readiness across extended sequences without excessive computational burden.

The paper contrasts direct and feature-based odometry paradigms, underscoring direct approaches' reliance on raw image intensity data, which, although non-convex, benefit significantly from the reliable short-term motion constraints provided by IMUs. This synergistic integration achieves a robust initial initialization, overcoming challenges akin to undetermined scale in monocular VIO. The paper discusses initialization through incorporation of scale and gravity as optimization variables, deploying an iterative refinement process that dynamically adjusts based on observed scale deviations, further bolstered by a robust marginalization strategy that mitigates drift over prolonged sequences.

The VI-DSO method holds substantial implications for autonomous systems that require immediate initialization and high-precision motion estimation. The joint estimation of scale, gravity, and other variables in a direct optimization framework paves the way for further refinements in real-time applications, notably in domains such as autonomous navigation and robotics.

Future prospects involve extending the framework to accommodate diverse sensor setups, enhancing the adaptability of the dynamic marginalization strategy, and exploring the integration with SLAM systems to leverage loop closure techniques potentially. As such, this paper contributes significantly to the ongoing development of robust visual-inertial systems capable of coping with real-world operational demands. The innovations presented substantiate the potential for enhanced real-time performance in dynamic environments, attending to the complexities of multi-sensor fusion and nonlinear optimization in odometry tasks.